Autoregressive Co-training

The implementation of the paper:

Autoregressive Co-Training for Learning Discrete Speech Representations
Sung-Lin Yeh, Hao Tang

Dependencies

pip install -r requirements.txt

Models

The co-training model described in the paper is defined in cotraining.py. Different components of the model are modular and can be easily modified.

Data

Data are processed to Kaldi I/O form, which uses scp files to map utterance ids to positions in ark files. Functions used to process .scp and .ark files can be found under dataflow/. We provide a data sample in sample/ for users to run the pipeline. Users can simply pluge in your custom dataloader here.

Train

python3 train.py --config config/cotraining.yaml

Pre-trained Models

Hours	Num codes	Model	dev93 (PER)	eval92 (PER)	Link
360	256	3-layer lstm with Marginalization	19.5	19.0	link
960	256	3-layer lstm with Marginalization	18.2	17.8	link